HMMER User's Guide

| Dept. of Genetics | WashU | Medical School | Sequencing Center | CGM | IBC| | Eddy lab | Internal (lab only) | HMMER | PFAM | tRNAscan-SE | Software | Publications |

Primary changes from HMMER 1.x

HMMER 2 is an almost complete rewrite of the original 1992-1996 HMMER code. A list of the major changes follows.

• [Plan7] The model architecture is changed. The ``Plan 7'' architecture^2.4 allows detailed control over parameters of local, global, and multidomain alignments. In Plan 7, you choose the alignment style when you build the model with hmmbuild, not when you search. A single database search program hmmsearch replaces the four search programs in HMMER 1.x.

• [Pfam support] There is much better support for the Pfam HMM database. HMMs have names; the HMM file format allows libraries of multiple HMMs per file; and there is a search program hmmpfam for searching a single query sequence against an HMM library. HMM databases like Pfam can be indexed for rapid access using hmmindex, and single HMMs can be retrieved from an HMM database using hmmfetch.

• [E-values] HMMER now reports both log-odds scores and a E-value. E-values can detect significant matches even when the log-odds score is negative. A new program hmmcalibrate empirically determines the necessary parameters for estimating E-values as best as possible.

• [Output] Database searches now sort and postprocess their results for prettier and more useful output. Output styles are deliberately similar to BLAST, to try to simplify the design of postprocessors.

• [Save file format] HMMER save files are now in an ASCII format which is easy for humans to read, portable across all architectures without byteswapping issues, and compact.

• [Decent defaults] The ``best'' average behavior of HMMER is now its default behavior. For example, mixture Dirichlet priors are now the default for protein model building. In HMMER 1.x, you had to invoke a number of ``experimental'' options to get better performance - this led to a bunch of really frustrating benchmarking in the literature, in which people compared fully optimized program X to the default behavior of HMMER 1, which was the 1994 basic Krogh/Haussler behavior rather than its actual power. (However, there's still a lot of bad benchmarking, because people don't understand how to benchmark profile software yet.)